Tabular Foundation Models Will Reshape Financial Services

SAP announced its acquisition of Prior Labs this week. While Prior Labs is not a financial services company, and SAP's move is not framed explicitly in that context, we believe this announcement foreshadows something fundamental about how the use of AI will evolve in financial services in the years ahead. The next generation of tools that financial institutions adopt may not look like "FinTech" at all.

Rethinking Where Value Will Be Created with AI

We have spent the last few years evaluating how AI will impact financial services, much of it focused on the application layer that serves financial services. Vertically-focused applications in financial services have been defensible historically. Regulation, licensing, distribution, and trust all contribute to durable moats, and there is strong reason to believe that many of these companies will continue to succeed.

At the same time, some of the most important systems in financial services are the horizontal platforms at the middleware and model infrastructure layers. Importantly, companies building deeper in the AI stack may not be focused exclusively on financial services, but we believe they will have an outsized impact on how financial institutions operate.

AI Is Not New to Financial Services

AI is already foundational to core workflows such as underwriting, fraud detection, transaction monitoring, and risk modeling. These systems have been powered for decades by machine learning (ML) models trained on tabular data. They are deeply embedded in how decisions are made across the industry.

The infrastructure behind these systems, however, reflects an earlier generation of ML.

Traditional ML models have been built one at a time for specific use cases. They require significant resources to develop, deploy, and maintain. Large institutions often run thousands of models, each taking on the order of weeks to months to develop, with meaningful backlogs and ongoing operational overhead. As demand for predictive systems continues to increase, only a new paradigm can sufficiently augment constrained ML teams.

The question is whether the existing model infrastructure approach can evolve. We believe the answer is no. Despite their rapidly accelerating maturity, LLM-based solutions are unlikely to be the answer for these highly structured, high-stakes workflows.

So does SAP. In a release announcing the Prior Labs acquisition, company CTO Philipp Herzig said:

"Early on, SAP recognized that the greatest untapped opportunity in enterprise AI wasn't large language models; it was AI built for the structured data that runs the world's businesses."

The New Paradigm of Model Infrastructure

A new class of models, often referred to as tabular foundation models (TFMs), is emerging to address this challenge.

These models are designed specifically for structured data and can generalize across a wide range of tasks with minimal additional tuning. Rather than building a separate model for each use case, organizations could begin to rely on pre-trained generalized models that can be applied across many problems.

It is important to be clear about what these models are and what they are not. They are not LLMs for spreadsheets. The mechanism by which LLMs process tabular data highlights this difference. LLMs interpret data sequentially and assign meaning to the order and proximity of values, as they would in language. This works well for text, but breaks down for structured data, where column order is arbitrary. Changing the order of two columns does not change the data, but an LLM will treat it as meaningfully different.

TFMs are built on fundamentally different assumptions and are optimized for statistical reasoning over structured data, making them suited to the types of problems that sit at the core of financial services.

A number of companies are actively building in the TFM space, including the aforementioned Prior Labs, (now part of SAP), Fundamental, Kumo, NeuralK, and Amazon's Mitra. These players are differentiated along dimensions such as data scale, ability to handle relational data, and enterprise readiness, but all are contributing to the development of this new paradigm of model infrastructure layer. TFMs are no longer a purely academic area of research.

Why TFMs Matter

The implications of this shift are meaningful.

From an operational perspective, these models can reduce the time required to deploy predictive systems, with workflows that previously took weeks or months now taking hours. This means reducing the number of people required to maintain old models and accelerating the productivity of data scientists building new models.

What LLMs have been to software engineers, TFMs may be to ML engineers. In the same way that LLMs abstracted away complexity in writing and debugging code, TFMs have the potential to abstract away much of the manual effort involved in building and maintaining predictive models.

As predictive models become easier to deploy, new use cases begin to emerge. This includes areas that have historically been constrained by data limitations or operational complexity. For example, the ability to underwrite underrepresented populations with limited historical data can expand the credit box. Similarly, payments companies may be able to launch fraud detection systems for new products, such as stablecoin payment risk, more quickly even without extensive prior datasets, rapidly accelerating timelines to bring products to market.
The opportunity here is to improve existing systems and to enable entirely new forms of decisioning.

What Will Matter in Financial Services

As the TFM category develops, the factors that determine success in financial services are likely to be distinct relative to other sectors.

Model performance will remain important. Improvements in underwriting accuracy can reduce losses, more effective fraud detection can cut chargeback losses, and stronger transaction monitoring more effectively curb money laundering. Latency will also matter in certain contexts, particularly in real-time decisioning environments.

However, the most under-appreciated factor may be explainability.

Financial institutions operate in highly regulated environments where decisions must be clear, auditable, and justifiable. Models outputting decisions that cannot meet these requirements are ineligible for high-stakes contexts such as credit decisions, regardless of their potential accuracy gains. The ability to interpret how a model arrives at its outputs is essential for adoption.

The most significant impact will come when these models are recognized by regulators as reliable systems for core workflows impacting consumers. Adoption into these workflows will depend not only on what these models can do, but on whether regulators and policymakers are comfortable with how they work. Building that comfort requires transparency, governance, and the ability to demonstrate that models behave in predictable and auditable ways.

This is an area where ecosystem engagement matters. At Flourish, we spend time at the intersection of policy and innovation through our work with partners such as FinRegLab and the Alliance for Innovative Regulation, as well as through hosting AI Roundtables that bring policymakers and builders into the same room. These conversations are critical to shaping how new technologies are understood and ultimately adopted within financial services.

In this context, the companies that succeed may not simply be those that build the most accurate models, but those that architect their infrastructure to fit regulatory requirements from day 1.

A Signal Worth Paying Attention To

TFMs represent a meaningful evolution in how predictive systems are built and deployed. SAP's acquisition of Prior Labs suggests that large enterprises are beginning to recognize the importance of this layer of model infrastructure in their move to leverage AI in core workflows.

For financial services, where decisioning sits at the core of the industry, this evolution has the potential to reshape how institutions operate. Not all of that change will happen at the application layer. Some of it will occur deeper in the stack, in the systems that power how decisions are made.

If you are building a TFM or building within this broader infrastructure and middleware layer in ways that could have a direct impact on financial services, we would welcome the opportunity to connect. We're paying close attention to this category.